SADiff: Coronary Artery Segmentation in CT Angiography Using Spatial Attention and Diffusion Model

Area of Science:

• Medical Imaging
• Cardiovascular Disease Research
• Artificial Intelligence in Healthcare

Background:

• Coronary artery disease (CAD) is a leading global cause of death, necessitating precise diagnostic tools.
• Accurate segmentation of coronary arteries in CT angiography (CTA) is vital for CAD diagnosis and treatment planning.
• Existing segmentation methods struggle with challenges like small vessel size, morphological variations, low contrast, and motion artifacts.

Purpose of the Study:

• To develop an advanced segmentation framework, SADiff, for improved coronary artery analysis from CTA images.
• To address limitations of conventional and early deep learning methods in segmenting complex coronary artery structures.
• To enhance the accuracy and generalizability of coronary artery segmentation for better CAD management.

Main Methods:

• Proposed SADiff, a hybrid framework integrating a dilated attention network (DAN) for region extraction.
• Incorporated a diffusion-based subnet for noise reduction in low-contrast areas.
• Utilized a striped attention network (SAN) for refining tubular structures with morphological variations.

Main Results:

• Achieved a Dice score of 83.48% and Hausdorff distance of 19.43 mm on the ImageCAS dataset, outperforming U-Net3D by 6.57% in Dice score.
• Demonstrated generalizability with a Dice score of 79.42% on the private ImageLaPP dataset.
• SADiff shows superior performance in segmenting coronary arteries compared to existing methods.

Conclusions:

• SADiff offers a more efficient and versatile approach to coronary artery segmentation from CTA.
• The framework shows significant potential for improving the accuracy of CAD diagnosis and treatment.
• The study highlights the effectiveness of the hybrid attention and diffusion-based approach for complex medical image segmentation.